On Friday, November 25, 2016 4:51:03 PM CET Vitaly Wool wrote:
> On Fri, Nov 25, 2016 at 9:41 AM, Arnd Bergmann <arnd@xxxxxxxx> wrote:
> > On Friday, November 25, 2016 8:38:25 AM CET Vitaly Wool wrote:
> >> >> diff --git a/mm/z3fold.c b/mm/z3fold.c
> >> >> index e282ba073e77..66ac7a7dc934 100644
> >> >> --- a/mm/z3fold.c
> >> >> +++ b/mm/z3fold.c
> >> >> @@ -884,7 +884,7 @@ static int __init init_z3fold(void)
> >> >> {
> >> >> /* Fail the initialization if z3fold header won't fit in one chunk */
> >> >> if (sizeof(struct z3fold_header) > ZHDR_SIZE_ALIGNED) {
> >> >> - pr_err("z3fold: z3fold_header size (%d) is bigger than "
> >> >> + pr_err("z3fold: z3fold_header size (%zd) is bigger than "
> >> >> "the chunk size (%d), can't proceed\n",
> >> >> sizeof(struct z3fold_header) , ZHDR_SIZE_ALIGNED);
> >> >> return -E2BIG;
> >> >
> >> > The embedded "z3fold: " prefix here should be removed
> >> > as there's a pr_fmt that also adds it.
> >> >
> >> > The test looks like it should be a BUILD_BUG_ON rather
> >> > than any runtime test too.
> >>
> >> It used to be BUILD_BUG_ON but we deliberately changed that because
> >> sizeof(spinlock_t) gets bloated in debug builds, so it just won't
> >> build with default CHUNK_SIZE.
> >
> > Could this be improved by making the CHUNK_SIZE bigger depending on
> > the debug options?
>
> I don't see how silently enforcing a suboptimal configuration is
> better than failing the initialization (so that you can adjust
> CHUNK_SIZE yourself). I can add something descriptive to
> Documentation/vm/z3fold.txt for that matter.
Failing at runtime when you know it's broken at compile-time
seems wrong, too. If you can't use z3fold with spinlock debugging,
you may as well hide the option in Kconfig based on the other ones.
Printing a runtime warning for the suboptimal configuration but
making it work anyway is probably better than just failing.
> > Alternatively, how about using a bit_spin_lock instead of raw_spin_lock?
> > That would guarantee a fixed size for the lock and make z3fold_header
> > always 24 bytes (on 32-bit architectures) or 40 bytes
> > (on 64-bit architectures). You could even play some tricks with the
> > first_num field to make it fit in the same word as the lock and make the
> > structure fit into 32 bytes if you care about that.
>
> That is interesting. Actually I can have that bit in page->private and
> then I don't need to handle headless pages in a special way, that
> sounds appealing. However, there is a warning about bit_spin_lock
> performance penalty. Do you know how big it is?
No idea, sorry. On x86, test_and_set_bit() seems to be only
one instruction to test/set the bit, followed by a conditional
branch, compared to a cmpxchg() for the raw_spin_lock(), so the
fast path seems pretty much the same.
Arnd
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>